HeFDN: Heterogeneous graph adversarial generation model for Chinese medical insurance fraud detection
Xinhua Dong, Hui Zhang, Zhigang Xu, Hongmu Han, Lifeng Jiang, Jintong Liu · Expert Systems with Applications · 2025
In recent years, the prevalence of medical insurance fraud has significantly impacted the national medical insurance fund. Furthermore, the task of medical insurance fraud detection faces considerable challenges due to the extreme imbalance of positive and negative samples. In practice, health insurance records contain longitudinal and cross-sectional multimodal data. Existing fraud detectors often disregard the multimodal information from heterogeneous neighbors (e.g., time, department, and treatment) and overlook the interaction details among multi-class data entities in claims behavior. To address these challenges, this study proposes the Heterogeneous graph-based Fraud Detection Network(HeFDN) for medical insurance fraud detection, which is based on heterogeneous graphs. The model represents the interaction behaviors of multi-class entities within medical insurance data through heterogeneous graphs. It also introduces a heterogeneous graph target node generator, termed HeGA, to address the issue of node imbalance classification in heterogeneous graphs. Leveraging the heterogeneous graph data structure, the model learns the topological structure and characteristics of actual minority class nodes by generating a set of virtual minority class target nodes. This approach aims to balance the data across different categories, thereby facilitating the distinction of minority class data in high-dimensional space. To validate the model’s effectiveness, this study conducted multiple experiments utilizing a public dataset and a dataset constructed from three real medical insurance sources from hospitals across various levels in Xianning City in China. Experimental results indicate that the proposed model enhances the AUC by an average of 18.2 % and the F1-Score by an average of 26.81 % compared to the most advanced algorithms reviewed.